CN104463249A - Remote sensing image airport detection method based on weak supervised learning frame - Google Patents

Abstract

The invention relates to a remote sensing image airport detection method based on a weak supervised learning frame. The method comprises the steps that firstly, the saliency of image blocks in a positive sample, the similarity of image blocks in a positive sample set and the inter-class difference between the positive sample and a negative sample are obtained, then, the Bayesian frame is utilized for fusing the three classes of information to obtain an initial positive and negative training set, then, iteration training of the training set is utilized for obtaining a final stable airport detector, the airport detector is used for detecting an airport of a tested image, and finally the airport detection result with better accuracy and robustness is obtained.

Description

A kind of remote sensing images airfield detection method based on Weakly supervised learning framework

Technical field

The invention belongs to computer vision algorithms make research field, relate to a kind of remote sensing images airfield detection method based on Weakly supervised learning framework, in remote sensing image data storehouse, accurately, robustly can detect the airport of image.

Background technology

The develop rapidly of remote sensing technology impels many satellites and airborne sensor can provide the optical imagery with high spatial resolution, and these images are widely used, as hazard management, and the reallocation of land, monitoring and traffic programme.In such applications, the automatic detection of nature or culture is a background task, and has attracted increasing research interest.The abundant spatial information comprised in the airport of remote sensing image and detailed structural information, provide new opportunity for we solve this challenging task.

Airfield detection in early days for remote sensing images is all unsupervised mode, first starts by and obtains area-of-interest to the cluster of pixel, and then Shape-based interpolation and spectrum information detect airport.After this, the method of many supervised learnings is proposed, they utilize the prior imformation in training sample effectively to learn out airfield detection model, by depending on a large amount of artificial mark training samples, then problem is converted into classification problem to go to solve, the method of supervised learning can obtain better effect than unsupervised method, and therefore, substantially all airfield detection systems are all based on supervised learning instrument.

The latest developments of remote sensing technology result in the explosive growth of the quality and quantity of satellite and aviation image.It also to bring in remote sensing image airfield detection task two day by day serious problems simultaneously.First, airfield detection method based on supervised learning needs a large amount of training datas usually, these training datas need manually by picture each airport be labeled in a rectangle frame, but, the manual mark of Large Scale Graphs image set costs dearly usually, sometimes or even insecure.Such as, for natural forms such as landslides, suitable manual mark needs a large amount of professional knowledge usually.In addition, for the culture such as aircraft and automobile, artificial mark is also difficult, because the coverage of these airport objects seems very little, when particularly comprising complicated texture in the background of image.Thus realizing accurately mark in these zonules is very difficult.In addition, if airport is blocked or pretends, the accuracy manually manually marked and reliability also can reduce.Therefore, for large-scale optical satellite and aerial image data collection, utilize Weakly supervised method training airplane field detector to become a noticeable direction.

Summary of the invention

The technical matters solved

In order to avoid the deficiencies in the prior art part, the present invention proposes a kind of remote sensing images airfield detection method based on Weakly supervised learning framework, solves and manually marks the problem wasted time and energy.

Technical scheme

A kind of remote sensing images airfield detection method based on Weakly supervised learning framework, it is characterized in that: utilize square moving window to extract segment in the remote sensing images of shooting, then utilize and train the airfield detection device obtained to classify to segment, and utilize the method for non-maxima suppression to keep score the highest window to solve the problem that lower of different scale gets window possibility high superposed, thus obtain final airfield detection result; The training step of described airfield detection device is as follows:

Step 1, carry out positive and negative sample classification to several remote sensing images randomly drawed in remote sensing image data storehouse, using the remote sensing images containing Airport information as positive sample image, the remote sensing images not containing Airport information are as negative sample image;

Step 2, utilize multiple dimensioned square moving window to extract segment in each width remote sensing images, the segment obtained by positive sample image is as the segment in positive sample set the segment that the segment that negative sample image obtains is concentrated as negative sample described sliding step is 1/3 of selected window size;

Step 3, calculate segment in positive sample set conspicuousness: using positive sample graph concentrate each segment around segment as dictionary, sparse coding χ is carried out to this segment _p≈ Dic _pα _p, according to the degree of rarefication of coding || α _p|| ₀with the residual error r that coding produces _p=χ _p-Dic _pα _p, obtain segment relative to the saliency value of ambient background wherein: χ _pfor segment original RGB feature, Dic _pand α _psegment respectively the sparse coefficient that the dictionary that around segment is formed produces with coding, represent the probability occurred;

Step 4, calculate the class inherited of positive and negative sample set: with reflection segment appear at probability that negative sample concentrates as the otherness between this segment and negative sample, and $\mathrm{Pr}\left({\left[f_p^{+}\right]}_j\right|y_p^{+}=0)=\underset{k=1}{\overset{K_j^{-}}{\mathrm{\Σ}}}{\mathrm{\π}}_{\mathrm{jk}}^{-}N\left({\left[f_p^{+}\right]}_j\right|{\mathrm{\μ}}_{\mathrm{jk}}^{-},{\mathrm{\σ}}_{\mathrm{jk}}^{2-}),$ Wherein for negative sample collection X ^-the weight of the kth gaussian component corresponding to middle jth dimensional feature, average and variance, represent the number of gaussian component in this gauss hybrid models;

Step 5, the approximate initial positive training set of acquisition: according to $\mathrm{Pr}(y_p^{+}=1|x_p^{+})\∝\frac{1}{\mathrm{Pr}\left(x_p^{+}\right)}[1-\mathrm{Pr}\left(x_p^{+}\right|y_p^{+}=0)\mathrm{Pr}(y_p^{+}=0)]$ Judge segment for the probability on airport, work as segment it is the probability on airport when being greater than threshold value, by this segment as the positive segment that initial training is concentrated, described threshold range is 0.5 ~ 1; Using the segment selected as approximate positive training set;

Step 6, calculate segment in positive sample image with the similarity of approximate positive training set: first utilize gauss hybrid models matching to be similar to sample distribution in positive training set; With as segment and the similarity between approximate positive training set; Wherein: represent segment jth dimensional feature, H represents segment feature total dimension, $\mathrm{Pr}\left({\left[f_p^{+}\right]}_j\right|y_p^{+}=1)=\underset{k=1}{\overset{K_j^{+}}{\mathrm{\Σ}}}{\mathrm{\π}}_{\mathrm{jk}}^{+}N\left({\left[f_p^{+}\right]}_j\right|{\mathrm{\μ}}_{\mathrm{jk}}^{+},{\mathrm{\σ}}_{\mathrm{jk}}^{2+}),$ Wherein for positive sample set the weight of jth dimensional feature in the gauss hybrid models that a middle kth segment is corresponding, average and variance;

Step 7, generation initial training collection: in the conspicuousness utilizing Bayesian frame to obtain step 2-6, class, similarity, class inherited merge, and finally obtain segment it is the probability on airport $\mathrm{Pr}(y_p^{+}=1|x_p^{+})=\frac{1}{\mathrm{Pr}\left(x_p^{+}\right)}[\mathrm{Pr}\left(x_p^{+}\right|y_p^{+}=1)\mathrm{Pr}(y_p^{+}=1)-\mathrm{Pr}\left(x_p^{+}\right|y_p^{+}=0)\mathrm{Pr}(y_p^{+}=0)],$ Wherein with be respectively segment x _pbelong to the probability of approximate positive training set and negative training set, calculate with nearest neighbor method;

Work as segment it is the probability on airport when being greater than threshold value, this segment as the positive segment that initial training is concentrated; Described threshold range is 0.5 ~ 1;

The negative training set that initial training is concentrated is produced by negative sample image stochastic sampling;

Step 8, training airplane field detector: utilize initial training set pair support vector machine to carry out repetitive exercise and obtain stable airfield detection device, in each iteration, a front iteration upgrade the up-to-date training set that obtains for training current airfield detection device, the current airfield detection device obtained after recycling training upgrades training set and is used as the training set of next iteration, until iteration terminates when this model starts to drift about, and using the previous airfield detection device of drift as final airfield detection device.

Described multiple dimensioned square moving window size is size ∈ { 60,100,130}.

Described nearest neighbor method calculates: segment x _pand the characteristic distance in positive and negative sample set between the immediate segment of characteristic distance is as segment x _pbelong to the probability of positive and negative sample set, formula is respectively $\mathrm{Pr}(x_p^{+}=1)=\exp \{-{\left|\right|x_p^{+}-\mathrm{Np}\left(x_p^{+}\right)\left|\right|}_1\},$ Wherein || || ₁represent L ₁norm, with represent segment x respectively _pthe feature of arest neighbors segment in positive and negative sample set.

Described remote sensing image data storehouse adopts Landsat storehouse.

Sparse coding in described step 3 adopts paper PJ.Han, P.Zhou, D.Zhang, G.Cheng, L.Guo, Z.Liu, S.Bu, and J.Wu, " Efficient, simultaneous detection of multi-class geospatial targetsbased on visual saliency modeling and discriminative learning of sparse coding; " ISPRS J.Photogramm.Remote Sens., vol.89, pp.37-48, the method in 2014.

Gauss hybrid models in described step 4 and 6 adopts paper C.M.Bishop, Pattern recognition andmachine learning.springer, the method in Aug.2006.

Described non-maxima suppression adopts paper G.Cheng, J.Han, L.Guo, X.Qian, P.Zhou, X.Yao, andX.Hu, " Object detection in remote sensing imagery using a discriminatively trainedmixture model; " ISPRS J.Photogramm.Remote Sens., vol.85, pp.32-43, the method in 2013.

Beneficial effect

A kind of remote sensing images airfield detection method based on Weakly supervised learning framework that the present invention proposes, first the conspicuousness of segment in positive sample is obtained, the similarity of segment in positive sample set, positive negative sample class inherited, then Bayesian frame is utilized to merge this three category information, obtain initial positive and negative training set, then these training set repetitive exercise are utilized to obtain final stable airfield detection device, the airport of test picture is detected with this airfield detection device, finally obtain and have more accuracy, the airfield detection result of robustness.

In the present invention, Weakly supervised study only needs to know in picture with or without airport, and do not need the particular location and the size that mark airport, therefore mark workload to greatly reduce, then the pictures with weak label obtained are utilized to train an airport detector, and according to training the airport detector obtained to carry out automatic marking again to picture, the label of training set is upgraded subsequently according to the result of automatic marking, after this this process of iteration, until detect that model starts drift, thus determine final airfield detection device, and utilize this airfield detection device to obtain to have more accuracy, the airfield detection result of robustness.

Accompanying drawing explanation

Fig. 1: the basic flow sheet of the inventive method

Fig. 2: experimental result picture

Fig. 3: PR result figure

Embodiment

Now in conjunction with the embodiments, the invention will be further described for accompanying drawing:

Hardware environment for implementing is: Intel Pentium 2.93GHz CPU computing machine, 2.0GB internal memory, the software environment of operation is: Matlab R2011b and Windows XP.All images that experiment have chosen in Landsat storehouse are as test data, and comprising 180 width shortwave-infrared remote sensing images in this database, is the database for testing remote sensing images airfield detection computation model of International Publication.

The present invention is specifically implemented as follows:

1, carry out positive and negative sample classification to several remote sensing images randomly drawed in remote sensing image data storehouse, using the remote sensing images containing Airport information as positive sample image, the remote sensing images not containing Airport information are as negative sample image;

2, utilize multiple dimensioned square moving window to extract segment in each width remote sensing images, the segment obtained by positive sample image is as the segment in positive sample set the segment that the segment that negative sample image obtains is concentrated as negative sample described sliding step is 1/3 of selected window size;

3, segment in positive sample set is calculated conspicuousness: using positive sample graph concentrate each segment around segment as dictionary, sparse coding χ is carried out to this segment _p≈ Dic _pα _p, according to the degree of rarefication of coding || α _p|| ₀with the residual error r that coding produces _p=χ _p-Dic _pα _p, obtain segment relative to the saliency value of ambient background wherein: χ _pfor segment original RGB feature, Dic _pand α _psegment respectively the sparse coefficient that the dictionary that around segment is formed produces with coding, represent the probability occurred;

4, the class inherited of positive and negative sample set is calculated: with reflection segment appear at probability that negative sample concentrates as the otherness between this segment and negative sample, $\mathrm{Pr}\left(x_p^{+}\right|y_p^{+}=0)=\underset{j=1}{\overset{H_2}{\mathrm{\Π}}}\mathrm{Pr}\left({\left[f_p^{+}\right]}_j\right|y_p^{+}=0),$ And

$\mathrm{Pr}\left({\left[f_p^{+}\right]}_j\right|y_p^{+}=0)=\underset{k=1}{\overset{K_j^{-}}{\mathrm{\Σ}}}{\mathrm{\π}}_{\mathrm{jk}}^{-}N\left({\left[f_p^{+}\right]}_j\right|{\mathrm{\μ}}_{\mathrm{jk}}^{-},{\mathrm{\σ}}_{\mathrm{jk}}^{2-}),$

Wherein for negative sample collection X ^-the weight of the kth gaussian component corresponding to middle jth dimensional feature, average and variance, represent the number of gaussian component in this gauss hybrid models;

5, approximate initial positive training set is obtained, with $\mathrm{Pr}(y_p^{+}=1|x_p^{+})\∝\frac{1}{\mathrm{Pr}\left(x_p^{+}\right)}[1-\mathrm{Pr}\left(x_p^{+}\right|y_p^{+}=0)\mathrm{Pr}(y_p^{+}=0)]$ Judge segment for the probability on airport, work as segment it is the probability on airport when being greater than threshold value, by this segment as the positive segment that initial training is concentrated, described threshold range is 0.5 ~ 1; Using the segment selected as approximate positive training set;

6, segment in positive sample image is calculated with the similarity of approximate positive training set: first utilize gauss hybrid models matching to be similar to sample distribution in positive training set; With as segment and the similarity between approximate positive training set; Wherein: represent segment jth dimensional feature, H represents segment feature total dimension, $\mathrm{Pr}\left({\left[f_p^{+}\right]}_j\right|y_p^{+}=1)=\underset{k=1}{\overset{K_j^{+}}{\mathrm{\Σ}}}{\mathrm{\π}}_{\mathrm{jk}}^{+}N\left({\left[f_p^{+}\right]}_j\right|{\mathrm{\μ}}_{\mathrm{jk}}^{+},{\mathrm{\σ}}_{\mathrm{jk}}^{2+}),$ Wherein for positive sample set the weight of jth dimensional feature in the gauss hybrid models that a middle kth segment is corresponding, average and variance;

7, initial training collection is produced: in the conspicuousness utilizing Bayesian frame to obtain step 2-6, class, similarity, class inherited merge, and finally obtain segment it is the probability on airport $\mathrm{Pr}(y_p^{+}=1|x_p^{+})=\frac{1}{\mathrm{Pr}\left(x_p^{+}\right)}[\mathrm{Pr}\left(x_p^{+}\right|y_p^{+}=1)\mathrm{Pr}(y_p^{+}=1)-\mathrm{Pr}\left(x_p^{+}\right|y_p^{+}=0)\mathrm{Pr}(y_p^{+}=0)],$ Wherein with be respectively segment x _pbelong to the probability of approximate positive training set and negative training set, calculate with nearest neighbor method;

Work as segment it is the probability on airport when being greater than threshold value, this segment as the positive segment that initial training is concentrated; Described threshold range is 0.5 ~ 1;

The negative training set that initial training is concentrated is produced by negative sample image stochastic sampling;

8, training airplane field detector: utilize initial training set pair support vector machine to carry out repetitive exercise and obtain stable airfield detection device, in each iteration, a front iteration upgrade the up-to-date training set that obtains for training current airfield detection device, the current airfield detection device obtained after recycling training upgrades training set and is used as the training set of next iteration, until iteration terminates when this model starts to drift about, and using the previous airfield detection device of drift as final airfield detection device.

The present invention selects PR curve to assess testing result.This curve definitions is under segmentation threshold change, the variation relation of accuracy rate (PRE) and recall rate (TPR).Computing formula is as follows:

$\mathrm{PRE}=\frac{\mathrm{TP}}{\mathrm{FP}+\mathrm{TP}}$

$\mathrm{TPR}=\frac{\mathrm{TP}}{P}$

Wherein FP is the false-alarm region detected; TP is the real police region territory detected, P is the region on airport in ground truth.Accompanying drawing 2 is some experimental results of the present invention, and accompanying drawing 3 is the PR curve of the inventive method.Can find out that the method that the present invention proposes can the more accurate detection with robustly realize airport to remote sensing images from experimental result.

Claims (7)

1. the remote sensing images airfield detection method based on Weakly supervised learning framework, it is characterized in that: utilize square moving window to extract segment in the remote sensing images of shooting, then utilize and train the airfield detection device obtained to classify to segment, and utilize the method for non-maxima suppression to keep score the highest window to solve the problem that lower of different scale gets window possibility high superposed, thus obtain final airfield detection result; The training step of described airfield detection device is as follows:

Step 1, carry out positive and negative sample classification to several remote sensing images randomly drawed in remote sensing image data storehouse, using the remote sensing images containing Airport information as positive sample image, the remote sensing images not containing Airport information are as negative sample image;

Step 2, utilize multiple dimensioned square moving window to extract segment in each width remote sensing images, the segment obtained by positive sample image is as the segment in positive sample set the segment that the segment that negative sample image obtains is concentrated as negative sample described sliding step is 1/3 of selected window size;

Step 3, calculate segment in positive sample set conspicuousness: using positive sample graph concentrate each segment around segment as dictionary, sparse coding χ is carried out to this segment _p≈ Dic _pα _p, according to the degree of rarefication of coding || α _p|| ₀with the residual error r that coding produces _p=χ _p-Dic _pα _p, obtain segment relative to the saliency value of ambient background || r _p|| ₁; Wherein: χ _pfor segment original RGB feature, Dic _pand α _psegment respectively the sparse coefficient that the dictionary that around segment is formed produces with coding, represent the probability occurred;

Step 4, calculate the class inherited of positive and negative sample set: with reflection segment appear at probability that negative sample concentrates as the otherness between this segment and negative sample, $\mathrm{Pr}\left(x_p^{+}\right|y_p^{+}=0)=\underset{j=1}{\overset{H_2}{\mathrm{\Π}}}\mathrm{Pr}\left({\left[f_p^{+}\right]}_j\right|y_p^{+}=0),$ And $\mathrm{Pr}\left({\left[f_p^{+}\right]}_j\right|y_p^{+}=0)=\underset{k=1}{\overset{K_j^{-}}{\mathrm{\Σ}}}{\mathrm{\π}}_{\mathrm{jk}}^{-}N\left({\left[f_p^{+}\right]}_j\right|{\mathrm{\μ}}_{\mathrm{jk}}^{-},{\mathrm{\σ}}_{\mathrm{jk}}^{2-}),$ Wherein for negative sample collection X ^-the weight of the kth gaussian component corresponding to middle jth dimensional feature, average and variance, represent the number of gaussian component in this gauss hybrid models;

Step 5, the approximate initial positive training set of acquisition: according to $\mathrm{Pr}(y_p^{+}=1|x_p^{+})\∝\frac{1}{\mathrm{Pr}\left(x_p^{+}\right)}[1-\mathrm{Pr}\left(x_p^{+}\right|y_p^{+}=0)\mathrm{Pr}(y_p^{+}=0)]$ Judge segment for the probability on airport, work as segment it is the probability on airport when being greater than threshold value, by this segment as the positive segment that initial training is concentrated, described threshold range is 0.5 ~ 1; Using the segment selected as approximate positive training set;

Step 6, calculate segment in positive sample image with the similarity of approximate positive training set: first utilize gauss hybrid models matching to be similar to sample distribution in positive training set; With as segment and the similarity between approximate positive training set; Wherein: represent segment jth dimensional feature, H represents segment feature total dimension, $\mathrm{Pr}\left({\left[f_p^{+}\right]}_j\right|y_p^{+}=1)=\underset{k=1}{\overset{K_j^{+}}{\mathrm{\Σ}}}{\mathrm{\π}}_{\mathrm{jk}}^{+}N\left({\left[f_p^{+}\right]}_j\right|{\mathrm{\μ}}_{\mathrm{jk}}^{+},{\mathrm{\σ}}_{\mathrm{jk}}^{2+}),$ Wherein for positive sample set the weight of jth dimensional feature in the gauss hybrid models that a middle kth segment is corresponding, average and variance;

Step 7, generation initial training collection: in the conspicuousness utilizing Bayesian frame to obtain step 2-6, class, similarity, class inherited merge, and finally obtain segment it is the probability on airport $\mathrm{Pr}(y_p^{+}=1|x_p^{+})=\frac{1}{\mathrm{Pr}\left(x_p^{+}\right)}[\mathrm{Pr}\left(x_p^{+}\right|y_p^{+}=1)\mathrm{Pr}(y_p^{+}=1)-\mathrm{Pr}\left(x_p^{+}\right|x_p^{+}=0)\mathrm{Pr}(x_p^{+}=0)],$ Wherein P be respectively segment x _pbelong to the probability of approximate positive training set and negative training set, calculate with nearest neighbor method;

Work as segment it is the probability on airport when being greater than threshold value, this segment as the positive segment that initial training is concentrated; Described threshold range is 0.5 ~ 1;

The negative training set that initial training is concentrated is produced by negative sample image stochastic sampling;

Step 8, training airplane field detector: utilize initial training set pair support vector machine to carry out repetitive exercise and obtain stable airfield detection device, in each iteration, a front iteration upgrade the up-to-date training set that obtains for training current airfield detection device, the current airfield detection device obtained after recycling training upgrades training set and is used as the training set of next iteration, until iteration terminates when this model starts to drift about, and using the previous airfield detection device of drift as final airfield detection device.

2. according to claim 1 based on the remote sensing images airfield detection method of Weakly supervised learning framework, it is characterized in that: described multiple dimensioned square moving window size is size ∈ { 60,100,130}.

3. according to claim 1 based on the remote sensing images airfield detection method of Weakly supervised learning framework, it is characterized in that: described nearest neighbor method calculates: segment x _pand the characteristic distance in positive and negative sample set between the immediate segment of characteristic distance is as segment x _pbelong to the probability of positive and negative sample set, formula is respectively $\mathrm{Pr}(y_p^{+}=0)=\exp \{-{\left|\right|x_p^{+}-\mathrm{Nn}\left(x_p^{+}\right)\left|\right|}_1\},\mathrm{Pr}(y_p^{+}=1)=\exp \{-{\left|\right|x_p^{+}-\mathrm{Np}\left(x_p^{+}\right)\left|\right|}_1\},$ Wherein || || ₁represent L ₁norm, with represent segment x respectively _pthe feature of arest neighbors segment in positive and negative sample set.

4. according to claim 1 based on the remote sensing images airfield detection method of Weakly supervised learning framework, it is characterized in that: described remote sensing image data storehouse adopts Landsat storehouse.

5. according to claim 1 based on the remote sensing images airfield detection method of Weakly supervised learning framework, it is characterized in that: the sparse coding in described step 3 adopts paper PJ.Han, P.Zhou, D.Zhang, G.Cheng, L.Guo, Z.Liu, S.Bu, and J.Wu, " Efficient, simultaneous detection of multi-class geospatialtargets based on visual saliency modeling and discriminative learning of sparse coding, " ISPRS J.Photogramm.Remote Sens., vol.89, pp.37-48, method in 2014.

6. according to claim 1 based on the remote sensing images airfield detection method of Weakly supervised learning framework, it is characterized in that: the gauss hybrid models in described step 4 and 6 adopts paper C.M.Bishop, Pattern recognition andmachine learning.springer, the method in Aug.2006.

7. according to claim 1 based on the remote sensing images airfield detection method of Weakly supervised learning framework, it is characterized in that: described non-maxima suppression adopts paper G.Cheng, J.Han, L.Guo, X.Qian, P.Zhou, X.Yao, andX.Hu, " Object detection in remote sensing imagery using a discriminatively trainedmixture model, " ISPRS J.Photogramm.Remote Sens., vol.85, pp.32-43, the method in 2013.